When is a PC an AI PC? Nobody seems to know or wants to tell Execs at the biggest PC makers seem to be salivating at the prospect of AI computers inflating their margins – even if they are unable or unwilling to define the emerging category. Lenovo pledged at the Canalys EMEA Forum 2023 to bring the AI form factor to market sometime between the second half of next year and early 2025, …
"An AI PC," he added, "is a PC that learns about you continuously, it's a PC that is your personal foundation model within the data within the PC and it is a PC that will be able to interact with you more naturally."
What you're actually saying is that Google can go cry in a corner, you're making a PC that will report our every click, hesitation and decision to some remote data store to be combed over incessantly by multiple scripts that will inevitably sell all that profiling to all and sundry.
Well I'm not having any of that. I've been using computers since the IBM PC. I don't need you to "help".
ah, but it won't report our every click, hesitation and decision to some remote data store to be combed over incessantly, it will process all those clicks, hesitations and decisions and it will send a regular, neat report to its masters. And you'll pay for electricity while it compiles those reports on your machine.
Bullshit? Yes, breath deeply, there's still a lingering fishy aroma to those that recall...
https://www.theregister.com/2015/02/19/superfish_lenovo_spyware/
I expect some of that extra processing will come in handy to obfuscate what they get up to
My definition would be a computer that can run a GPT3-equivalent LLM locally, since GPT3 is the one that kick-started the current AI revolution. This would need at least 64 to 128GB of RAM accessible by a GPU or NPU. Thus the Mac Studio qualifies, as would a PC with a couple of nVidia 4070s. Today's price for such a config is about $5000, but expect that to fall quickly.
> Today's price for such a config is about $5000, but expect that to fall quickly
So, approximately ten times more expensive than a perfectly serviceable PC. And all that oomph will be eaten up running the LLM, need to add on some more to allow it to do something useful as well.
Given the current tech climate, just how long is it going to take to drop to a price-point that it could actually be considered worth buying by the general office or home user? Have you noticed nVidia (or similar GPU providers) dropping its prices recently?
Not to mention the power used by a rig like you describe!
It won't though, the only constraining factor for *running* an LLM on your local machine right now is RAM...and RAM is cheap, especially in older machines.
I work in a lab where quite a few people run their own LLMs on their machines to assist with software development, and their machines can be anywhere up to 10 years old (because they have specific hardware that needs to be compatible).
If you're training an LLM, then it's a different story.
Why all the fuss about the price? The people who are pushing Machine Learning (it's not their fault their spell checker changed it to Artificial Intelligence - they would never over hype anything) are EXPERTS!
I have been following them and a number of evangelists, gurus and influencers on TiK-Dork and idiotstagram for ages, and thanks to their advice I can EASILY afford to but a PC to do this. I will simply sell some of my NFTs, get paid in craptocurrency and as the blockchain is secure I will be able to withdraw my funds which have grown incredibly thanks to Sam Bankcard-Fraud and the team at FTX.
You luddites need to get with the program (or is that programme, when I have an AI PC it will put the right word..or Pogrom, or Pomegranate, not sure)
Exactly! It's amazing how many commentards seem to think that AI is a passing fad and even more astonishing is how many of these commentards have no idea how LLMs work. Are we sure this is still a tech forum?
I've got quantised AI models working on Raspberry Pi's...granted, they're not as broadly powerful as say GPT4...but for the specific use cases the models were trained for, they exceed GPT4 in at least one area.
I took an existing, "good enough", LLM and performed some additional training with it to make it competent at wifi surveys, network analysis and device fingerprinting...it's a 1.5GB quantised model that runs pretty well on a Raspberry Pi Zero 2W that I built into a handheld device that I carry around for doing network surveys and audits with...it isn't lightning fast, but that doesn't really matter because no matter how fast your LLM runs, the quality of the results will be the same...also, you don't need instant results...you just need the speed of the results to be faster than a human...which they are, no matter what speed they run at...because for a human to complete a basic report on a network based on the collected data, it would take at least a day...my Pi based LLM will have a basic report for me within 5-10 minutes...using that report, I can bang out my own report (after carefully reading the AI one, just in case, the AI hasn't provided any barking mad results yet, but you always have to be vigilant, as with any person / machine you farm out a task to) in about an hour.
What the hell are you on about?
I have a local LLM running on my network that I use daily, it runs on a quad core Xeon-E3 from 9 years ago with 32GB RAM...the performance of the LLM is roughly equivalent to GPT3.5...the results are probably better, because I am using a models tailored to my use cases.
AI models are computationally intensive to train, but trivial to use...especially quantised models.
Ironically, the Mac silicon based Macs are absolutely dismal at AI currently compared to stuff much older than itself...especially generative AI for making images.
It's a PC whose manufacturers have decided they can make lots more money by associating it with the latest technological bandwagon, and who will then work backwards from that to figure out how to justify the tag, most likely by slapping some privacy-invading AI crapware on.
With the growing use of virtualization, VDI, cloud storage and cloud computing over the last 10-15 years, virtually all programs and data of the PC user are increasingly resident on the cloud. As a result, a PC has virtually become a Dumb Terminal these days. I wonder what user-specific data is even there on the local hard disk of the PC anymore for AI PC to get trained on and work as a Personal Assistant.
I'm sure you could offload your counterstrike playing to an AI that has learned to "slot enemies" far better than you ever could, thanks to a few billion rounds of training courtesy of users of 'cloud gaming platforms' like the one Microsoft just bribedvigorously lobbied regulators around the world to allow them to purchase and slurp.
And I'm sure a further iteration same AI definitely won't be used by a robot to 'slot' you and/or your children / grandchildren one day.
In short...yes...
I've experimented with using an HDMI capture card passing the video through various computer vision algorithms. It is entirely possible to have a second monitor setup on a second machine, running computer vision to identify targets in Counter Strike...there is a small delay, but it doesn't really impact anything if your intention is to spot enemies that you otherwise might not have seen.
I haven't yet tried anything out on CS2...but it's only a matter of updating the dataset to reflect the new graphics.
I haven't experimented with sending inputs to the "host" machine yet, but there is no reason it couldn't be done and therefore there is no reason you couldn't build an AI to play Counter Strike for you.
The real battle is bringing the latency down.
This does actually sound like it could be a benefit. Instead of spending two hours composing an email or presentation filled with lies, half-truths and errors, a manager could have the LLM compose an email or presentation filled with lies, half-truths and errors in only a couple of minutes.
This will enable a sixtyfold increase in the quantity of pointless timewasting, which will obviously lead to a sixtyfold increase in profits.
There is an awful lot of marketing bollocks about but this is impressive by anyone’s standards.
As for having a PC that monitors what you do all the time, why would anyone want that? And as AI doesn’t actually exist but is being used as a synonym for a LLM, why would i want a PC that runs a LLM? To do what?
Utter bollocks
why would i want a PC that runs a LLM?
A software co might want to run an in-house server for private CoPilot-like service.
Or an electronics manufacturer wants an in-house service for visual product inspection - they don't want to leak their proprietary design or quality data.
Or a retailer wants to use it for visually identifying goods at a self check out, etc, which might require continual updating for new products.
In house also means it won't fail due to internet interruption - one less snafu.
No one's rooting for a PC that "learns everything about me" and most likely sends the results back to Lenovo HQ.
This is merely a ploy to inflate their margins by jumping on the AI Hype Bandwagon. I predict it will have a simple Neural Processor on-board or NN GPU software that don't cost much more to add but can be sold for premium prices. Or so they think.
> the prospect of AI computers inflating their margins – even if they are unable or unwilling to define the emerging category.
Yup. Primarily a marketing term to convince those with too much money that they are getting something that will inflate their egos abilities.
So make something that is a little bit better, even if only in entirely irrelevant specifications, and then flog it as revolutionary. At three times the price.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
